Jay Hack Coins 'Lecun-Marcus Dodge' on LLM Code Limits
Jay Hack coins term amid debate on autoregressive LLMs for code with tools.
Gary Marcus posted screenshots of Yann LeCun statements and claimed LeCun has become his clone after four years. Jay Hack replied to LeCun and others that critics now call the shift the Lecun-Marcus dodge. Hack asked what the same people would have said in June 2025 about ten-hour programming sessions using autoregressive LLMs plus linters and compilers. Susan Zhang posted that some rewrite history to claim they always supported systems built around models. Aidan Clark replied that the figures show no interest in accountability for past words. Other replies referenced von Neumann lectures on reliable systems from unreliable parts.
@ylecun @gabriberton They are calling this the "Lecun-Marcus dodge" If we had asked in June 2025 wether 10h programming with linters + compilers hooked up to autoregressive LLMs would be possible, what would you have said?
Combined views
Reactions from ranked influencers
16 poststhis is how these people somehow manage to rewrite history to Always Be On The Right Side as if, you know, the possibility of Using Them in a System was previously never on the table or anything brother literally fundraised to invent something new, and it's going to be…
🚨 Yann LeCun’s four year long transformation into becoming Gary Marcus is now complete!
🚨 Four no doubt painful years later, @YannLeCun’s transformation into becoming a perfect clone of @GaryMarcus is now complete! 🚨
@suchenzang Honestly, there was a time not so long ago when scaling-pilled people would argue against using LLMs in a system on bitter lesson grounds. I don’t think that’s a strawman
@suchenzang The “pretraining thesis” of Gwern’s Scaling Hypothesis strikes me as a canonical example. The vibe there is that optimizing the final few bits of cross entropy is the way to build an intelligence system, not taking a well-pretrained base model as input to a larger…
@suchenzang This isn’t to dunk on that essay btw; it’s a nice read! I just think it’s useful context for the range of views at the time
